Gym

OpenAI's way to create a reinforcement learning environment, used by most RL implementations.

• TODO: Further my understanding of OpenAI's Gym

Gym seeks to create standard environments, enabling better benchmarks to drive research in reinfrocement learning.

An environment contains:

• action_space: space object corresponding to valid actions

• observation_space: space object corresponding to valid observations

• reward_range: corresponding min and max possible rewards

A step in an environment contains:

• observation (object): pixel data, angles, velocities or board states

• reward (float): amount of reward achieved by previous action

• done (boolean): whether to reset the environment, termination action.

• info (dict): diagnositic information (i.e. raw probabilities for last state)

The framework of reinforcement learning

• How to create custom environment: https://gym.openai.com/docs/#environments

• Keras-RL uses gym style environments (although the included ones are blank abstract classes)